Image contrast statistics influence population receptive fields in lateral occipital cortex

(1)

Image contrast statistics influence population

receptive fields in lateral occipital cortex

Alban Voppel

December 13, 2016

Abstract

Previous research has shown that scene and object recognition in hu-man visual perception makes use of image contrast distributions present in natural scenes. We performed a population receptive field (pRF) ex-periment using stimuli that varied in their image contrast distributions. Characteristics of population receptive fields measured using fMRI in lat-eral occipital areas were found to be influenced by contrast stimuli present in background as well as in bar-pass stimuli. This points to use of these image contrast statistics in visual processing.

1 Introduction

1.1 Image statistics

In everyday life, complex and variable natural scenes are interpreted and pro-cessed quickly by the visual system. Efficient scene recognition is necessary to visually process sensory input in a timely manner, but how the brain does this is an unresolved issue. In the search for a fast and efficient way to classify and process highly variable natural visual scenes, various statistical regularities, no-tably in the distribution of local contrast, have been observed (Brunswik and Kamiya, 1953; Geusebroek et al., 2002). A Weibull function describes the shape of the distributions of contrast strengths present in natural images (Simoncelli, 1999; Geusebroek et al., 2002). When a Weibull function is fit to this distribu-tion, it has two free parameters, representing the scale (β) and the shape (γ) of the contrast distribution (see Equation 1). These image statistics can be used to classify natural scenes (Scholte et al., 2009; Groen et al., 2013).

Natural scenes images have similarly been classified using the contrast en-ergy (CE) and spatial coherence (SC) statistical measures. CE is an approxi-mation of β while SC approximates γ; both measures meaningfully modulated responses in single trial ERP signal amplitudes when participants where asked to rate scenes on naturalness. When images are ordered based on these values, a quick categorization of images can be made with complex natural scenes at one extreme, and simple and clear figure-on-uniform-backgrounds on the other (see Figure 1) (Scholte et al., 2009; Groen et al., 2012, 2013).

Scholte et al. showed that an artificial neural model based on the β and γ parameters of visual scenes accurately explains ERP differences. Participants were shown natural images; the accompanying β and γ parameters were used as explanatory variables in explaining differences in ERP responses. EEG data was measured over multiple subjects; the Weibull β and γ parameters explained up to 50% variance of the ERP signal of electrode lz, overlying visual cortex, between 80 and 200 ms after stimulus presentation (Scholte et al., 2009).

Further research into the possible use of these measures by the brain has been performed. Across-subject EEG data recorded while participants looked at natural scenes, combined with the Weibull parameters of these images, has been used to train an artificial neural model. This model predicted subject EEG responses to 100 novel natural scenes. The model was then presented with novel EEG data of a presented natural scene; the model had to classify, based on the Weibull parameters, which scene the subject had seen. Performance of the model was 90%. Notably, in misclassification of natural scenes the images were visually similar to correct images in their spatial layout and contrast values. This is an argument for the thesis that the Weibull parameters capture information used by the human visual processing system, since they convey information regarding the neural response (Ghebreab et al., 2009).

Dead Leaves occlusion stimuli (Hsiao and Millane, 2005) have been used to replicate these findings. In this stimuli creation paradigm, image elements

(3)

Figure 1: Distribution of natural scenes, ordered by CE (an approximation of β) and SC (an approximation of γ). Four representative images are shown. Figure from Groen et al., 2013.

composed of varied sized-disks are placed randomly on an image (see Figure 2 for representative examples). Varying amounts of overlap and edge contrasts in the image elements used result in images which differ in their β and γ values. Performance of the classifier was 94% correct using the Dead Leaves stimuli, an argument for the hypothesis that the distribution of contrasts and not semantic or conceptual information is what is captured and used in visual processing of natural scenes (Ghebreab et al., 2009).

Thus, β (or CE) and γ (or SC) are presumed to be ecologically relevant parameters that are used in the visual system for rapid classification of natural scenes and images. However, how these parameters influence neural processing is not yet clear (Scholte et al., 2009; Groen et al., 2013).

1.2 The BOLD response and Population Receptive Fields

we record the average activations of the whole population of neurons in one voxel, i.e. the population receptive field In parts of the cortex associated with visual processing, neurons have a receptive field, which defines to which region in visual space they are sensitive and will respond to stimuli. Due to the nature of functional magnetic resonance imagining (fMRI) we record the average activity-related blood-oxygen level dependent signal (BOLD) of the whole population of neurons in one voxel, i.e. the population receptive field (pRF). Using specifically designed fMRI stimuli, the size, position in visual space, as well as eccentricity from the retina can be estimated for the receptive field of each voxel (Dumoulin and Wandell, 2008).

(4)

The method introduced by Dumoulin and Wandell for estimating the pRF goes further than previously used ring-and-wedge patterns, allowing estimation not only of pRF location but also of size (Dumoulin and Wandell, 2008). In this method, bars moving across the screen from different directions are used to elicit BOLD activity in voxels in the visual system. Using computational models, the per-voxel measured BOLD response timecourse can be converted into a receptive field for that voxel, modeled as a two-dimensional Gaussian in retinotopic space, with an amplitude representing percent signal change of BOLD activity, a size of the pRF in degrees of visual arc and an eccentricity of the pRF as the distance relative to the fovea (Dumoulin and Wandell, 2008).

Earlier research has shown that the lateral occipital (LO) area in the ventral visual stream is especially sensitive to the γ Weibull parameter; BOLD activity here is well described by this parameter (Scholte et al., 2013). Here, we examined this effect by measuring pRF parameters in area LO for conditions differing in β and γ contrast parameters present in mapping stimuli and background. We used the Dumoulin and Wandell pRF mapping method, with stimuli as well as backgrounds composed of Dead Leaves image elements to explore the effect of image contrast statistics on pRFs, thus examining a proposed fast and efficient mechanism for visual scene perception.

2 Methods

2.1 Subjects

Four subjects (1 female, ages from 24 to 28) participated in the study. All participants had normal or corrected-to-normal vision and gave their informed consent. The experiments were approved by the ethics committee of the faculty of social and behavioral sciences at the University of Amsterdam. Participants received e 40 after completing both sessions of the experiment.

2.2 MRI acquisition

All MRI data was acquired on a Philips Achieva XT 3T scanner (Philips Medical Systems) at the Spinoza Center Amsterdam, equipped with a 32-channel head coil. Each subject completed two sessions of approximately one hour each. 12 scans were completed per session; one whole-brain T1 weighted scan with 1 mm isotropic voxels, one high-resolution T2 weighted with the same box as the functional scans, and ten randomly presented functional runs (voxel size 2.5 x 2.5 x 2.75 mm, TR = 1000 ms), with each run corresponding to one condition. Each functional run took approximately six minutes and was composed of 12 trials, with 4 trials being blank and 8 being bar-passes. The trials were presented in a semi-random fashion to prevent the 4 blank trials occurring after each other.

2.3 Eye tracking and physiology

Eye tracking (9-point calibrated EyeLink 1000, SR systems, Ontario, Canada) was used to track fixation efficiency as well as eye drift. Physiological data (heart rate and breathing rhythm) was measured for use in preprocessing.

(5)

Figure 2: Examples of bars created using the Dead Leaves method in varying β and γ parameter space. From top to bottom; Checkerboard used as control condition; condition with low β - low γ parameters, condition with low β - high γ parameters, condition with high β - low γ parameters, condition with high β - high γ parameters.

2.4 Stimuli

2.4.1 Weibull distribution

The distribution of contrast strengths in images can be captured and modeled by the Weibull distribution; see equation 1.

p(f ) = c e(

f −µβ

)

γ

(1) Here c is a normalization constant that transforms the distribution of con-trast frequencies into a probability distribution. The scale parameter β describes the width of the contrast histogram. It varies with the amount of variation in contrasts present in the image. The parameter γ describes the shape of the histogram, and varies with the amount of scene clutter present.

The µ parameter represents the origin of the distribution. Its position is influenced by uneven illumination. To achieve illumination invariance, the µ parameter was normalized out, leaving β and γ as free parameters (Scholte et al., 2009; Ghebreab et al., 2009).

2.4.2 Dead Leaves stimuli generation

The Dead Leaves stimuli method is a method for creating artificial images de-void of semantic content. Images are created by placing geometric shapes with varying size, edges, and illumination randomly on a canvas. In the present study, only circular shapes were used (see Figures 2 and 3). By varying parameters determining size, overlap as well as hues of these circles, a varied collection of images can be created, ensuring a wide range in Weibull β-γ space (Ghebreab et al., 2009).

2.5 Stimulus presentation

Visual stimuli were created using MATLAB version R2014b (MathWorks) and the Psychophysics Toolbox (Brainard, 1997). 24 types of Dead Leaves image cat-egories were generated following (Ghebreab et al., 2009). For these images, we

(6)

(a) A bar with high β and low γ parameters passes diagonally over a neutral-gray background with a red fixation dot.

(b) In this condition, a bar with high β and γ parameters passes vertically over a high β and γ background.

Figure 3: Examples of Dead Leave generated stimuli bars passing the circular aperture, with fixation dot.

created contrast distributions by using a Gabor-wavelet pyramid filter (Scholte et al., 2009). From these contrast distributions, Weibull parameters β and γ were estimated using a maximum likelihood estimator following Ghebreab et al..

We selected four image categories from the resulting images; the 3rd per-centile for each parameter was taken for each of these conditions to ensure a wide spread in β-γ space, similar to natural scenes. Additionally, a standard blocked stimulus with maximum contrast was created as a control condition. From these selected images, 100 bars per condition were created to be used as bar-pass stimuli (see Figure 2).

Stimuli were presented on a 1920 x 1200 pixel beamed display with a circular aperture of 900 pixels, corresponding to 13.4 degree visual angle (dva). Subjects were instructed to fixate on a red dot in the center of the beamed display for all 12 trials per condition. This fixation dot (0.07 dva) changed color to black three or four times per trial for a length of 0.25 seconds per change. Subjects had to respond to this color change by pressing a button with their right index finger after each change. This ensured that the participants remained fixated and attended the center of the screen and not the bar pass.

During the 8 non-empty trials in each of the 10 conditions, a bar composed of Dead Leaves image elements (width 1.4 dva), refreshing to a new Dead Leave image of the same category 100 times per bar-pass moved completely across the circular aperture in 24 seconds, in one of 8 cardinal and diagonal directions. Thus, the refresh rate of the bar was effectively 4 Hz. The inter-trial time was three seconds.

Each of the five conditions (four different Dead Leave bars and one con-trol bar) was presented once with the circular aperture filled with a neutral-luminance gray background. Additionally, each of the five conditions was pre-sented with a randomly chosen static background composed of a Dead Leaves images with high β and γ Weibull parameters (See Figure 3b).

(7)

2.6 MRI preprocessing

For each subject, one out of two sessions was discarded because of missing runs, wrong slicebox planning or problems with the setup. For the remaining session per subject, the session T1-weighted image was automatically segmented using Freesurfer (Fischl, 2012). Per session, a single target functional run was selected, based on the amount of constant magnetic field (B0) warping present. This run was registered to the Freesurfer-segmented T1-weighted image using the bbregister function.

Subsequently, all functional runs were motion corrected to their own middle volume using the FSL software utility MCFLIRT (Jenkinson et al., 2012). Then, the functional runs were registered both linearly (FLIRT) and non-linearly (FNIRT) to the mean-motion corrected functional target run.

Low frequency drifts in BOLD signal were removed using a third order Savitzky-Golay filter (Savitzky and Golay, 1964) with a window length of 120 seconds, corresponding to a cutoff of 0.00833 Hz. Finally, the arbitrary BOLD units were changed to percent signal change, based on the values of BOLD units per functional run.

2.7 pRF data analysis

For pRF estimation, we used a procedure combining a data-driven approach (Lee et al., 2013) with a more standard model-based approach (Dumoulin and Wan-dell, 2008). First, a data-driven approach is used to provide a robust although coarse estimate of the center of the pRF. Then, the model-based approach re-fines this initial position estimate and also yields estimations of the pRF size and amplitude.

In the data-driven approach, a subject-specific Haemodynamic Response Function (HRF) convolved with a design matrix encoding the stimulus position for each TR of the run is used to perform a Ridge regression. This procedure does not explain signal variance, but serves to find the position of the pRF by searching for the maximum beta value (for a more in-depth explanation, see (Lee et al., 2013)). Subsequently, the position of maximal beta value is set as the starting value of a gradient-descent fitting procedure in the model-based approach, using Powell’s algorithm (Powell, 1964). For each voxel i the pRF P (i, x, y) was modeled as an isotropic Gaussian, defined in visual space (x, y):

P (i, x, y) = exp − ((x − xoi) 2_{+ (y − yo} i)2 2σ2 i ) (2)

Here, xo and yo reflect the center of the pRF as found by the data-driven approach. σ represents the pRF size. The stimulus design matrix S(i, t, x, y) was composed of a model of stimulus on each TR t with the bar present as 1 and absence as 0. Using a standard HRF h, the overlap between P (i, x, y) and S(i, t, x, y) was convolved, producing the BOLD prediction for each voxel g(i):

g(i) = βi+ αi(h ∗ (P · S)) (3)

Here α reflects the amplitude parameter and β the baseline for this voxel. Because there were only 8 bar passes per condition per subject we first performed a fit procedure on the data of all conditions concatenated together to achieve robust parameter estimates (the all-fit). From these estimates, one separate

(8)

Figure 4: The Freesurfer-inflated right cortical hemisphere of subject 3. Retino-topy based on polar angle coordinates and eccentricity was used to define func-tional areas in the visual stream. These coordinates were derived from the fit procedure on all conditions.

model-based fit procedure per condition (each with a separate xo, yo, σ and a shared β ) was performed, with the parameters of these fits initialized at the estimates of the all-fit. To ensure that we only selected visually responsive voxels, all voxels in the all-fit with a r2_{< 0.10 were discarded.}

2.8 Retinotopy

To determine functional areas in the brain, we performed retinotopic analysis for subjects. In the visual cortex, neurons are organized in retinotopic maps that can be used to define specific functional areas (Wandell et al., 2007). Two participants had participated in previous experiments; retinotopic maps were already available for these participants. For the two other subjects, functional areas were hand-defined based on eccentricity and polar angle mapping as de-rived from the all-fit fitting procedure (see Figure 4). We used the TkSurfer utility included in the Freesurfer package. For a review of retinotopy and maps in the visual cortex, see (Wandell et al., 2007).

2.9 Statistics

With functional areas defined for each subject, pRF data for voxels were pooled across subjects based on region of interest (ROI). For all statistical compar-isons, a bootstrap procedure was performed across all the voxels in the ROI. Outliers were removed if they crossed a threshold of three standard deviations. Additionally, we discarded all voxels with an estimated pRF size bigger than our stimuli aperture as these voxels cannot be reliably estimated.

(9)

Figure 5: pRF parameters for high versus low γ conditions in area LO. Left: Mean pRF size in area LO for high γ versus low γ conditions. The difference is not significant, p = 0.408. Right: Mean pRF amplitude in area LO for high γ versus low γ conditions. The difference is significant; p = 0.001. Black lines represent standard deviation in pRF parameters.

3 Results

We calculated per-voxel measures of pRF size, amplitude and eccentricity using a bar-pass pRF mapping procedure. In four conditions, bar stimuli were com-posed of image elements widely spread in β-γ parameter space, corresponding to contrast energy (CE) and spatial coherence (SC) measures present in nat-ural images. Additionally, a bar-pass condition composed of black and white squares with maximum contrast was presented. Each condition was repeated with a standard-luminance gray background and a background composed of Dead Leaves image elements high in β-γ space.

3.1 Effect of γ in area LO

From previous literature, the γ image parameter, corresponding to contrast en-ergy, was a good fit for activity in area LO (Scholte et al., 2013). We performed an estimation of pRF parameters in area LO for conditions high in γ space, as well one for conditions low in γ space. By comparing per-voxel pRF estimations in different conditions, we can estimate whether they significantly differ between conditions.

In conditions with a high γ parameter compared to conditions with a low γ parameter the size of the pRFs in area LO does not significantly change (p = 0.408). However, there is a significant difference in pRF amplitude between conditions differing in γ (p = 0.001)(Figure 5). We conclude that a higher γ parameter is associated with higher pRF amplitudes in area LO, but not significantly with a change in pRF size in area LO.

(10)

Figure 6: pRF parameters for high versus low β conditions in area LO. Left: Mean pRF size in area LO for high β versus low β conditions. The difference is significant, p < 0.001. Right: Mean pRF amplitude in area LO for high γ versus low γ conditions. The difference is significant; p > 0.001. Black lines represent the standard deviation in pRF parameters.

3.2 Effect of β in area LO

Similar to the fit procedure for the Weibull γ procedure, we performed an anal-ysis of the effect on pRF parameters in area LO for the β parameter, corre-sponding to scene coherence. In conditions with a high β parameter compared to conditions with a low β parameter we found a significant change in pRF amplitude (p < 0.001), Similar to the effect of γ in area LO. Conditions with a higher β parameter lead to higher pRF amplitudes in area LO (Figure 6). However, we additionally found an significant change regarding pRF size. pRF size was larger for low β conditions, and these values significantly differ from high β conditions (p < 0.001). A low β leads to bigger pRF sizes. This effect was not found for the γ parameter. Both β as well as γ parameters mainly seem to modulate pRF amplitude, but only β modulates pRF size in area LO.

3.3 Influence of contrast in background

We tested whether pRF size and amplitude in area LO were meaningfully changed by having a background high in β-γ space. This corresponds to a background that is richly filled with high-energy contrast, over which the bar moved. We compared this background to a standard-luminance neutral gray background.

We performed pRF fit procedures for all conditions, divided in two groups; all conditions with and all conditions without a background, both initialized at the values of the all-fit procedure. To differentiate between a global effect present throughout the visual pathway caused by presenting a background rich in contrasts or specific processing of Weibull contrast parameters in area LO we compared the primary visual cortex (V1) and area LO (Figure 7).

(11)

versus non-background (p < 0.001); pRF size was generally larger for neutral-gray background conditions. No significant difference in pRF amplitude was found (p = 0.14).

In area LO, similar to V1, there was a significant difference between gray background and β-γ-backgrounds visible in pRF size (p < 0.001), with neutral-gray backgrounds pRFs being larger. However, amplitude in area LO was sig-nificantly modulated by background, a result not found in V1. pRF ampli-tude recorded with a Dead Leave background high in β-γ space are lower than neutral-gray backgrounds (p < 0.001). Although the strength of pRF ampli-tude measured in percent of signal change in V1 was higher than in area LO, the difference between the conditions was significant in LO. Since this modu-lation was not present in V1, it points toward a specific sensitivity to Weibull image statistics in area LO, expressed in pRF amplitude. Natural image con-trasts high in β-γ space - corresponding to scene coherency and contrast energy - thus influence area LO specifically in the pRF amplitude.

4 Discussion

4.1 Image statistics and pRFs in area LO

We have noted effects of Weibull image statistics, corresponding to scene co-herency and contrast energy present in natural images, on pRF parameters in area LO. pRF Amplitude was affected by the γ Weibull parameter in area LO. We have shown that the β parameter also significantly influenced pRF am-plitude in area LO. Additionally, the β parameter influenced pRF size; the γ parameter did not. The earlier finding of BOLD activity modulation in LO by the Weibull γ parameter combined with these novel pRF effects point towards the use of these statistical contrast measures in area LO (Scholte et al., 2013).

Similarly, the difference between pRF parameters in V1 and area LO using conditions with a background are an argument for the fact that specific brain areas process contrast distributions in different ways. Since in natural scenes there is almost always a collection of contrast over the whole field of view, pRF estimation using a background will be more suited to explore scene perception as it approaches ecological conditions. Since the experiment did not look at the relative effect of other backgrounds with different β and γ parameters, a quantification of this background effect as it correlates to other image statistic parameters was not possible. Future research should take this into consideration. It is important to note that although we do note a significant difference between conditions differing in Weibull parameter space, it is unclear from our experiment alone whether this difference is actually meaningfully used in natural scene recognition. For instance, it might be that one of the known functions of area LO, object recognition (e.g. (Malach et al., 1995)), plays a confounding role in the responses currently seen. Since the current stimuli were composed of circular image elements, overlapping in different amounts and with different edge and contrast ratios, the process of recognizing these stimuli as ball- or circle-like objects might play a role. Differences between conditions in pRF estimation might be due to variability in this object-detection due to the aforementioned parameters.

(12)

(a) Mean pRF size in V1. Differ-ence = p < 0.001.

(b) Mean pRF amplitude in V1, in percent of BOLD signal change. The difference is not significant, p = 0.14.

(c) Mean pRF size in area LO, in degrees of visual angle. Difference = p < 0.001.

(d) Mean pRF amplitude in area LO, in percent of BOLD signal change. Difference = p < 0.001.

Figure 7: Difference in effect of high β-γ background versus neutral gray back-ground in V1 and LO. Top row: size and amplitude of pRFs in V1, as well as the difference between conditions. Bottom row: size and amplitude of pRFs in area LO. Conditions presented with a background high in β-γ space elicit a significantly lower pRF amplitude in area LO, yet do not do so in V1. Black lines represent standard deviation.

statistics in previous experiments, where explained variance and correlations between image statistics and natural scenes have been further explored. These experiments suggest a causal, information-rich role being played by the process-ing of contrast distribution parameters present in natural scenes (Scholte et al., 2009, 2010; Groen et al., 2012, 2013).

(13)

More research is needed regarding image statistics, both in object recognition with and without backgrounds present, as well as in scene perception. Here we have shown a significant interaction between image statistics and pRF parame-ters in area LO. A quantitative relationship between image statistic parameparame-ters and brain responses could be found, both in ERP research by presenting natural scenes, as well as in fMRI experiments.

4.2 Future directions and methodological improvements

To better characterize the influence of the image contrast parameters future research might investigate other points of interest with more bar-passes; for example, to repeat the experiment in either only β or γ Weibull contrast space - with the other parameter fixed - but with more discrete categories, to examine the exact modulation of these contrast parameters on pRF responses in various brain areas. This might lead to a quantitative assessment of pRF change per Weibull parameter.

Recently, new models for BOLD responses to various visual stimuli have been proposed (Kay et al., 2013a,b). Use of these models allows more robust predictions and fits on visual responses in fMRI response. Not only spatial summation, but also the two-stage cascade model provide better, more robust fits in pRF fitting procedures. Especially non-linear spatial summation effects of stimuli is relevant for the research conducted here, as the stimuli we used are composed of moving bars. Our experiment thus effectively presented, with every movement of the bar, non-completely filling stimuli to parts of the cortex. Neurons responding to their partially-filled receptive field for stimuli present in the bar pass will have reacted differently than modelled using the current tech-nique. Estimation of pRF characteristics, as well as the estimation of influence of Weibull parameters, would thus be improved by incorporating these models (Kay et al., 2013a,b).

Further research or validation studies should include a custom HRF per sub-ject Hs in equation 3 to improve fits. Since estimations of pRF parameters are

especially dependent on HRF, an experimental stimulus should be included that allows for precise HRF estimation per subject. Research has made a per-voxel custom HRF estimation possible. There is promising research of HRF estima-tion based on differing computaestima-tional models (Pedregosa et al., 2015). These methods can be used to further strengthen model estimations and robustness of predictions of pRF parameters (Dumoulin and Wandell, 2008).

4.3 Mishaps in current research

Because of procedural mishaps, for each subject one out of two sessions had to be discarded. Data would be strengthened by incorporating multiple sessions. Due to time constraints, we did not perform retroicor analysis (Glover et al., 2000) for subjects. Future research should incorporate these steps. Also, although we measured eye and gaze positions through eye tracking, fixation normalization should have taken place; this did not happen due to time constraints.

(14)

5 Conclusion

Concluding, we provide evidence that the β (Contrast energy) and γ (Scene co-herency) Weibull parameters play a meaningful role in visual perception. Esti-mated pRF amplitudes in area LO are significantly different between conditions, both for the β as well as γ Weibull parameter. pRF size is also influenced by the β parameter. From these effects it follows that information about natural scenes is at least partially encoded in contrast distributions of these scenes, and the visual system makes use of these measures.

6 References

Brainard, D. H. (1997). The psychophysics toolbox. Spatial vision, 10:433–436.

Brunswik, E. and Kamiya, J. (1953). Ecological cue-validity of’proximity’and of other gestalt factors. The American journal of psychology, 66(1):20–32.

Dumoulin, S. O. and Wandell, B. A. (2008). Population receptive field estimates in human visual cortex. Neuroimage, 39(2):647–660.

Fischl, B. (2012). Freesurfer. Neuroimage, 62(2):774–781.

Geusebroek, J.-M., Smeulders, A. W., et al. (2002). A physical explanation for natural image statistics.

Ghebreab, S., Scholte, S., Lamme, V., and Smeulders, A. (2009). A biologically plausible model for rapid natural scene identification. In Advances in Neural Information Processing Systems, pages 629–637.

Glover, G. H., Li, T.-Q., and Ress, D. (2000). Image-based method for retro-spective correction of physiological motion effects in fmri: Retroicor. Magnetic resonance in medicine, 44(1):162–167.

Groen, I. I., Ghebreab, S., Lamme, V. A., and Scholte, H. S. (2012). Spa-tially pooled contrast responses predict neural and perceptual similarity of naturalistic image categories. PLoS Comput Biol, 8(10):e1002726.

Groen, I. I., Ghebreab, S., Prins, H., Lamme, V. A., and Scholte, H. S. (2013). From image statistics to scene gist: evoked neural activity reveals transition from low-level natural image structure to scene category. The Journal of Neuroscience, 33(48):18814–18824.

Hsiao, W. and Millane, R. (2005). Effects of occlusion, edges, and scaling on the power spectra of natural images. JOSA A, 22(9):1789–1797.

Jenkinson, M., Beckmann, C. F., Behrens, T. E., Woolrich, M. W., and Smith, S. M. (2012). Fsl. Neuroimage, 62(2):782–790.

Kay, K. N., Winawer, J., Mezer, A., and Wandell, B. A. (2013a). Compres-sive spatial summation in human visual cortex. Journal of neurophysiology, 110(2):481–494.

(15)

Kay, K. N., Winawer, J., Rokem, A., Mezer, A., and Wandell, B. A. (2013b). A two-stage cascade model of bold responses in human visual cortex. PLoS Comput Biol, 9(5):e1003079.

Lee, S., Papanikolaou, A., Logothetis, N. K., Smirnakis, S. M., and Keliris, G. A. (2013). A new method for estimating population receptive field topography in visual cortex. Neuroimage, 81:144–157.

Malach, R., Reppas, J., Benson, R., Kwong, K., Jiang, H., Kennedy, W., Led-den, P., Brady, T., Rosen, B., and Tootell, R. (1995). Object-related activity revealed by functional magnetic resonance imaging in human occipital cortex. Proceedings of the National Academy of Sciences, 92(18):8135–8139.

Pedregosa, F., Eickenberg, M., Ciuciu, P., Thirion, B., and Gramfort, A. (2015). Data-driven hrf estimation for encoding and decoding models. NeuroImage, 104:209–220.

Powell, M. J. (1964). An efficient method for finding the minimum of a function of several variables without calculating derivatives. The computer journal, 7(2):155–162.

Savitzky, A. and Golay, M. J. (1964). Smoothing and differentiation of data by simplified least squares procedures. Analytical chemistry, 36(8):1627–1639.

Scholte, H. S., Ghebreab, S., Smeulders, A., and Lamme, V. (2010). Lateral occipital cortex responsive to correlation structure of natural images. Journal of Vision, 10(7):1363–1363.

Scholte, H. S., Ghebreab, S., Waldorp, L., Smeulders, A. W., and Lamme, V. A. (2009). Brain responses strongly correlate with weibull image statistics when processing natural images. Journal of Vision, 9(4):29–29.

Scholte, H. S., Sligte, I., Groen, I., Lamme, V., and Ghebreab, S. (2013). The posterior part of the lateral occipital complex analyzes the spatial correlation structure of natural visual scenes. Journal of Vision, 13(9):1098–1098.

Simoncelli, E. P. (1999). Modeling the joint statistics of images in the wavelet domain. In SPIE’s International Symposium on Optical Science, Engineering, and Instrumentation, pages 188–195. International Society for Optics and Photonics.

Wandell, B. A., Dumoulin, S. O., and Brewer, A. A. (2007). Visual field maps in human cortex. Neuron, 56(2):366–383.

Image contrast statistics influence population receptive fields in lateral occipital cortex